When Should a Molecular Taste Prediction Be Trusted? Evaluating the Reliability of Fingerprint-Based Taste Classification

Accurate molecular taste prediction is important for the development of drugs and the support of research in food science. Despite increasing use of machine learning to predict the taste of molecules, little attention has been paid to determining when those predictions should be trusted. We built and compared a machine learning model on a dataset of molecular fingerprints from FlavorDB against a baseline model, and then evaluated its ability to identify the taste of a molecule from molecular properties using performance metrics. The fingerprint-based model achieved the strongest performance with a macro-F1 score of 0.83, but the reliability varied between taste classes, with weaker performance in underrepresented tastes and in molecules that were less similar to those in the training data.

These findings indicate that machine learning can predict taste from molecular structure, but that performance metrics alone are insufficient for judging prediction reliability because reliability varies across taste classes, dataset balance, and a molecule’s similarity to the training data.